Load short-term probability prediction method based on multi-valued quantile regression and ensemble learning

Through the method based on multi-value quantile regression and ensemble learning, the problems of poor interpretability in load short-term probability prediction and low accuracy in multi-objective prediction are solved, and more accurate and comprehensive load short-term probability prediction is achieved, improving interval coverage comprehensiveness of rate and predicted results.

CN119009931APending Publication Date: 2024-11-22STATE GRID ZHEJIANG ELECTRIC POWER CO LTD SHAOXING POWER SUPPLY CO +2
View PDF 0 Cites 11 Cited by

Patent Information

Application Number
CN202410885517.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-03
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

In the prior art, there are problems such as poor explanatory and low multi-objective prediction accuracy. In particular, ordinary quantile regression can easily lead to crossing between multiple quantile prediction values, and the short-term probability of the entire load cannot be fully obtained. Condition.

Method used

Using a method based on multivalue quantile regression and ensemble learning, the historical load data is divided into historical benchmark data and historical predecessor data, an integrated learning sub-model is constructed, and an integrated learning initial model is constructed through initial dynamic weighting and weighting parameters updates. Combined with the adaptive bandwidth kernel density estimation calculation method, the continuous probability density curve is output to improve the accuracy and comprehensiveness of the prediction results.

Benefits of technology

Effectively improve the cross-over problem of multiple load quantile prediction values, improve the interval coverage accuracy of load interval prediction, reduce the interval prediction width, and provide more accurate and comprehensive short-term load probability prediction results, which facilitates load regulation strategies. Formulation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119009931A_ABST
    Figure CN119009931A_ABST
Patent Text Reader

Abstract

The invention discloses a load short-term probability prediction method based on multi-valued quantile regression and ensemble learning, and the method comprises the following steps: constructing a plurality of ensemble learning sub-models through the historical reference data, and obtaining a load quantile prediction value training set through the plurality of ensemble learning sub-models; constructing an integrated learning initial model by using the plurality of integrated learning sub-models; outputting a weight parameter update value based on the integrated learning model loss function to obtain an integrated learning model; outputting a load quantile prediction value set corresponding to the to-be-predicted time sequence according to the to-be-predicted time sequence data and the integrated learning model; and outputting a continuous probability density curve according to the load quantile predicted value set based on an adaptive bandwidth kernel density estimation algorithm to obtain a load short-term probability prediction result. The method has the beneficial effects that the problem of intersection of multiple load quantile predicted values is improved, the interval coverage rate accuracy of load interval prediction can be effectively improved, and the interval prediction width can be reduced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the technical field of load forecasting, and in particular to a method for short-term probability forecasting of load based on multi-value quantile regression and ensemble learning. Background Art

[0002] With the continuous advancement of the construction of new power systems, traditional deterministic load forecasting results have certain limitations in describing load uncertainty. In the relevant technology, there are studies on short-term probability forecasting of loads: the probability forecasting method based on parameter models makes parameter forecasts by assuming load distribution, and can directly give the probability density or cumulative distribution curve of the load based on the parameters output by the model, but it is also very dependent on the rationality of the distribution assumptions. In actual situations, it is difficult for the load at a certain moment to strictly obey a certain distribution.

[0003] Probabilistic forecasting formally includes probability parameter estimation, interval forecasting, quantile regression, etc. In interval forecasting, Xu Shihong et al. directly predicted the upper and lower limit intervals, and their forecast results can provide the extreme conditions of load fluctuations, but their forecast results are difficult to provide information on the probability density and probability distribution of the load.

[0004] At present, the quantile regression in the non-parametric model has the characteristics of simple operation, and various artificial intelligence algorithms can be used to establish a relatively reasonable model, but the following problems still exist: 1. Ordinary quantile regression easily leads to cross-over between multiple quantile prediction values, and the interpretability becomes poor; 2. Multi-objective prediction problems corresponding to multiple quantile load values.

[0005] The Chinese patent "A super short-term probabilistic wind power forecasting method based on spatiotemporal multi-scale and K-SDW", publication number: CN 115600500A, publication date: January 13, 2023, specifically discloses the following steps: S1, decomposing the normalized target wind speed into multiple subsequences through variational mode decomposition; S2, reconstructing the subsequences and adjacent spatial wind speed sequences into spatiotemporal candidate features, and performing spatiotemporal multi-scale feature selection; S3, using the quantile regression model to predict each subsequence. This scheme uses a sub-model to predict each component separately, and the overall prediction performance is poor, and it is impossible to fully obtain the short-term probability of the entire load. Summary of the invention

[0006] In order to solve the problems of poor interpretability and low multi-objective prediction accuracy in the prior art for short-term probability load prediction, the present application provides a short-term probability load prediction method based on multi-value quantile regression and ensemble learning, which divides historical load data into historical benchmark data and historical prior data, constructs an ensemble learning sub-model with the historical benchmark data, and then obtains a training set of load quantile prediction values ​​corresponding to the historical prior data, performs initial dynamic weighting on several ensemble learning sub-models, and constructs an ensemble learning initial model. At this time, the load quantile prediction value output by the ensemble learning initial model is the output value of the initial dynamic weighting performed on the load quantile prediction value training set Based on the loss function of the ensemble learning model and the actual load value in the historical prior data, the loss parameter of the load quantile prediction value output by the ensemble learning initial model compared to the actual load value is calculated to obtain the weight parameter update value. The weight coefficients of each sub-model in the ensemble learning initial model are updated with the weight parameter update value, so that the performance of the ensemble learning model is optimized and the accuracy of the final prediction result is improved. Then, the prediction result of the ensemble learning model is used to output a continuous probability density curve based on the adaptive bandwidth kernel density estimation algorithm to show the short-term probability prediction result of the load, improve the accuracy and comprehensiveness of the load prediction results in the future period, and facilitate the formulation of load control strategies. It can improve the crossover problem of multiple load quantile prediction values, effectively improve the accuracy of the interval coverage rate of the load interval prediction and reduce the interval prediction width, and obtain the short-term load interval prediction and probability density prediction results of the 10kV busbar of the distribution network.

[0007] To achieve the above technical purpose, the present application provides a technical solution, which is a load short-term probability forecasting method based on multi-value quantile regression and ensemble learning, comprising the following steps: S1: obtaining historical load data, using the historical load data to divide into historical benchmark data and historical prior data according to the time series, using the historical benchmark data to build a number of ensemble learning sub-models, and using the historical prior data and a number of ensemble learning sub-models to obtain a load quantile prediction value training set corresponding to the historical prior data; S2: performing initial dynamic weighting on a number of ensemble learning sub-models to construct an ensemble learning initial model; S3: outputting a weight parameter update value based on the ensemble learning model loss function according to the historical prior data and the load quantile prediction value training set, performing secondary dynamic weighting with the weight parameter update value, updating the ensemble learning initial model, and obtaining an ensemble learning model; S4: Obtain the time series data to be predicted, and use the time series data to be predicted and the integrated learning model to output the set of load quantile prediction values ​​corresponding to the time series to be predicted; S5: Based on the adaptive bandwidth kernel density estimation algorithm, a continuous probability density curve is output according to the set of load quantile prediction values ​​to obtain the short-term probability prediction result of the load.

[0008] Furthermore, the method of constructing a plurality of integrated learning sub-models using historical benchmark data also includes: randomly dividing the historical benchmark data into a plurality of training data subsets, and constructing a corresponding integrated learning sub-model with each training data subset to obtain a plurality of integrated learning sub-models.

[0009] Furthermore, the method of constructing a corresponding integrated learning sub-model with each training data subset also includes: constructing an integrated learning basic sub-model including a feature selection network, an LSTM encoding layer, an LSTM decoding layer and a gated residual network; using the feature selection network to identify influencing characteristic factors in the training data subset; using the LSTM encoding layer to obtain load short-term change characteristics based on the influencing characteristic factors; and using the load short-term change characteristics as input to the LSTM decoding layer and the gated residual network to construct an integrated learning sub-model.

[0010] Furthermore, the output of weight parameter update values ​​based on the ensemble learning model loss function according to historical prior data and load quantile prediction value training set also includes: constructing an ensemble learning model loss function using a multi-value quantile loss function and a maximum likelihood estimation sorting loss.

[0011] Furthermore, the method of constructing an integrated learning model loss function using a multi-valued quantile loss function and a maximum likelihood estimation ranking loss also includes: constructing a multi-valued quantile regression model based on a pinball loss function combined with quantile regression and a neural network model, and obtaining a multi-valued quantile regression loss function using the multi-valued quantile regression model.

[0012] Furthermore, the method of outputting a continuous probability density curve based on a set of load quantile prediction values ​​based on an adaptive bandwidth kernel density estimation algorithm also includes: calculating the local density under the initial bandwidth for each load quantile prediction value in the set of load quantile prediction values; calculating the geometric mean density with the local densities corresponding to all load quantile prediction values; calculating the adaptive bandwidth of each load quantile prediction value with the geometric mean density; and calculating the probability density estimate of the load under the time series to be predicted using the adaptive bandwidth, and outputting a continuous probability density curve.

[0013] Furthermore, the use of historical load data to divide into historical benchmark data and historical advance data according to time series includes: dividing the historical load data into a number of time segments according to preset time segments, using odd-numbered time segments as historical benchmark time segments, using even-numbered time segments as historical advance time segments, and obtaining historical benchmark data and historical advance data based on the historical benchmark time segments and the historical advance time segments.

[0014] Furthermore, S2 also includes: using a mixer to perform initial dynamic weighting on a plurality of ensemble learning sub-models to construct an ensemble learning initial model.

[0015] Furthermore, the multi-value quantile regression loss function is: Among them, ML(y i x i ) is the condition input x i Time i The multi-valued quantile regression loss function, y i is a vector of several load quantile forecast values ​​at time i, m is the total number of quantiles, q j represents the j-th quantile, The quantile is q j The loss function when , g(xiqj) is the conditional input xi, quantile q j Multi-value quantile prediction model for .

[0016] Furthermore, the historical load data at least includes historical meteorological data, historical time series data and historical load data.

[0017] The beneficial effects of this application are as follows: 1. By improving the loss function of multi-value quantile regression and adding sorting constraints to it, the occurrence of crossover between quantile values ​​is effectively reduced, and the interpretability of the load probability prediction model is improved; 2. For the multi-objective prediction problem of multiple load quantile values, an integrated learning prediction model is constructed, which improves the accuracy of the interval prediction of the load probability prediction results and the reliability of the probability prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of the short-term probability load forecasting method based on multi-value quantile regression and ensemble learning in this application.

[0019] Figure 2 This is a structural diagram of the integrated learning sub-model of the load short-term probability forecasting method based on multi-value quantile regression and integrated learning in this application. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solutions and advantages of the present application more clear, the present application is further described in detail below in conjunction with the drawings and examples. It should be understood that the specific implementation method described here is only an optimal embodiment of the present application, which is only used to explain the present application and does not limit the scope of protection of the present application. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0021] like Figure 1 As shown, as the first embodiment of the present application, for the short-term load interval prediction and probability density prediction of the 10kV bus of the distribution network, a load short-term probability prediction method based on multi-value quantile regression and ensemble learning includes the following steps: S1: Obtain historical load data, use the historical load data to divide it into historical benchmark data and historical prior data according to the time series, use the historical benchmark data to build several ensemble learning sub-models, and use the historical prior data and several ensemble learning sub-models to obtain the load quantile prediction value training set corresponding to the historical prior data; S2: Perform initial dynamic weighting on several ensemble learning sub-models to construct an ensemble learning initial model; S3: Based on the loss function of the ensemble learning model, the weight parameter update value is output according to the historical prior data and the load quantile forecast value training set, and the weight parameter update value is used to perform secondary dynamic weighting, update the ensemble learning initial model, and obtain the ensemble learning model; S4: Obtain the time series data to be predicted, and output a set of load quantile prediction values ​​corresponding to the time series to be predicted using the time series data to be predicted and the integrated learning model; S5: Based on the adaptive bandwidth kernel density estimation algorithm, a continuous probability density curve is output according to the load quantile prediction value set to obtain the load short-term probability prediction result.

[0022] In this embodiment, the historical load data is divided into historical benchmark data and historical prior data, and an integrated learning sub-model is constructed with the historical benchmark data, thereby obtaining a training set of load quantile prediction values ​​corresponding to the historical prior data, and performing initial dynamic weighting on several integrated learning sub-models to construct an integrated learning initial model. At this time, the load quantile prediction value output by the integrated learning initial model is the output value of the initial dynamic weighting performed on the load quantile prediction value training set. Based on the integrated learning model loss function and the actual load value in the historical prior data, the loss parameter of the load quantile prediction value output by the integrated learning initial model compared with the actual load value is calculated to obtain an updated weight parameter value, and the weight coefficient of each sub-model in the integrated learning initial model is updated with the updated weight parameter value, so that the performance of the integrated learning model is optimized and the accuracy of the final prediction result is improved. Then, the prediction result of the integrated learning model is used to output a continuous probability density curve based on the adaptive bandwidth kernel density estimation algorithm to present the short-term probability prediction result of the load, thereby improving the accuracy and comprehensiveness of the load prediction result for the future period and facilitating the formulation of load control strategies. It can improve the intersection problem of multiple load quantile prediction values, effectively improve the interval coverage accuracy of load interval prediction and reduce the interval prediction width, and obtain the short-term load interval prediction and probability density prediction results of the 10kV bus of the distribution network.

[0023] The historical load data at least includes historical meteorological data, historical time series data and historical load data. The historical load data is divided into historical benchmark data and historical leading data according to the time series, including: The historical load data is divided into several time segments according to the preset time segments, with odd-numbered time segments as historical benchmark time segments and even-numbered time segments as historical prior time segments, and historical benchmark data and historical prior data are obtained based on the historical benchmark time segments and the historical prior time segments.

[0024] The number of time series segments is an even number, that is, there is a corresponding historical preceding time series segment for each historical reference time series segment. It can be understood that for the historical reference time series segment, each historical preceding time series segment is a relative future time series of each historical reference time series segment.

[0025] In other cases, the historical load data is divided into a training set and a validation set, and the training set and the validation set are divided into historical benchmark data and historical prior data respectively. In the training process of the ensemble learning sub-model, the historical benchmark data and historical prior data in the training set are used as two parts of input to perform the training of the ensemble learning sub-model.

[0026] Using historical benchmark data to build several integrated learning sub-models also includes: The historical benchmark data is randomly divided into several training data subsets, and a corresponding ensemble learning sub-model is constructed with each training data subset to obtain several ensemble learning sub-models.

[0027] In this embodiment, several integrated learning sub-models use the same model architecture, but are trained using different training data subsets, and ultimately each integrated learning sub-model has different training weight parameters.

[0028] like Figure 2 As shown, m is an adjustable hyperparameter of the model, but it does not exceed the number of predictions of multi-valued quantiles. Constructing the corresponding ensemble learning sub-model with each training data subset also includes: Construct an integrated learning basic sub-model including feature selection network, LSTM encoding layer, LSTM decoding layer and gated residual network; Use feature selection networks to identify influential feature factors in a subset of training data; The LSTM encoding layer is used to obtain the short-term load change characteristics based on the influencing characteristic factors; An integrated learning sub-model is constructed using the short-term load variation characteristics as the input of the LSTM decoding layer and the gated residual network.

[0029] The integrated learning sub-model contains two parts of input. The first part of the input is the historical known feature vector, which is obtained through historical benchmark data, including historical meteorological factors, load levels, and holiday information; the other input is the known load feature value at the corresponding time, which is obtained through historical prior data, that is, the relative future meteorological forecast value, holiday information, etc. After the input historical time series feature vector is passed through the feature selection network, it can identify the features that have the greatest impact on load forecasting, that is, the influencing feature factors, and pass through the LSTM encoding layer as part of the input features of the decoding process. This process can provide short-term load change features as part of the input of the load forecasting decoding layer, and then use the gated residual network to extract features from the encoding vector. After the same operation with the future feature vector, the long-term load change information is extracted in the multi-head attention mechanism layer. In this process, the extraction of short-term change information and the long-term load change information are fused, and the two processes together constitute the fusion of long-term and short-term time series features in the sub-model. By fusing historical and future time series features, using technical means such as LSTM and multi-head attention mechanism, the prediction accuracy of future load is improved.

[0030] In this embodiment, the prediction results of multiple time series fusion sub-models are integrated by a mixer, and the prediction results are 99 quantile load prediction values ​​to obtain the optimal prediction result. Therefore, step S2 also includes: The mixer is used to perform initial dynamic weighting on several ensemble learning sub-models to construct an ensemble learning initial model.

[0031] The mixer adopts a dynamic weighted structure, with the prediction result of the kth ensemble learning sub-model as y k , the mixer performs dynamic weighting as: Where n is the number of sub-models in the ensemble learning model; is the j-th quantile q corresponding to time i j The load forecast value at time y k|i is the vector of the 99 quantile load forecast value combinations of the kth ensemble learning sub-model at time i; ω k The weight of the prediction result of the kth ensemble learning sub-model; y i It is the output of the mixer, i.e., a vector of several load quantile prediction values ​​at time i.

[0032] In this embodiment, each ensemble learning sub-model is set with a corresponding initial weight, which may be an average weight.

[0033] Then, the loss function of the ensemble learning model is constructed: the maximum likelihood estimation ranking loss is introduced into the multi-value quantile loss function to form the loss function used in the ensemble learning model training. At this time, based on the ensemble learning model loss function, the output weight parameter update value of the training set according to the historical prior data and the load quantile forecast value also includes: The loss function of the integrated learning model is constructed using the multi-value quantile loss function and the maximum likelihood estimation ranking loss.

[0034] The basis of multi-value quantile regression is traditional quantile regression. Through quantile regression, we can observe the relationship between the dependent variable and the independent variable at different quantiles and obtain the quantile regression process: in, is the load forecast value when the corresponding quantile is q at time i; F i Enter x for the condition i Time i Cumulative distribution function, that is, when the load y i The value is less than the quantile value The corresponding probability is is the inverse of the probability cumulative distribution function.

[0035] At this time, the ensemble learning model loss function is constructed with multi-value quantile loss function and maximum likelihood estimation sorting loss, and the loss function also includes: A multi-valued quantile regression model is constructed based on the pinball loss function combined with quantile regression and neural network model, and the multi-valued quantile regression loss function is obtained by the multi-valued quantile regression model.

[0036] Based on the pinball loss function, quantile regression is combined with the neural network model. Quantile regression is used as the neural network loss function to construct a neural network quantile regression model: Among them, L q is the loss function corresponding to the quantile q; f q (x) is the conditional quantile prediction model when the quantile value is q. PL is easy to optimize and is very effective when processing data with different quantiles. The neural network quantile regression model is constructed by training the network model to minimize the loss.

[0037] Traditional quantile regression can only construct conditional quantile regression under a single quantile value, while the multi-value quantile regression model uses a multi-output model to simultaneously predict multiple quantile values ​​of the load at a certain moment. The multi-value quantile regression model is constructed as follows: in, The corresponding quantile at time i is qj The load forecast value at time i; g(x) is the multi-value quantile forecasting model, which can simultaneously predict the m quantile values ​​of the load at time i.

[0038] At this time, the multi-value quantile regression loss function is: Among them, ML(y i x i ) is the condition input x i Time i The multi-value quantile regression loss function is the cumulative sum of multiple quantile losses and can be regarded as a multi-objective loss function with equal weights.

[0039] In the multi-value quantile regression problem, when the input feature is X at a certain moment, the m quantile prediction value vector of the load is recorded as Y, P XY is the unknown but fixed joint probability distribution of X and Y, the ranking constraint of multi-valued quantiles can be transformed into minimizing the expected ranking loss R(h): h(x i |q j )=sort(g(x i |q1),g(x i |q2),...,g(x i |q j )); Where l is the loss coefficient, P(x i ,y i ) is the joint probability density distribution, and h is the multi-quantile numerical ranking function.

[0040] Since there is a 0-1 loss and ranking function h(x i |q j ) makes it difficult to optimize the model. By constructing a substitution function Φ, a ranking loss function, that is, a negative logarithmic loss function, is constructed. The substitution function Φ is: At this time, the ranking loss function based on maximum likelihood estimation is: Order the multi-valued quantile values ​​output by the model by maximizing the probability: Then, the loss function of the integrated learning model is constructed using the multi-value quantile loss function and the maximum likelihood estimation ranking loss; Among them, L(y i,f(x i )) is the loss function of the integrated learning model, which includes two parts: multi-value quantile regression loss and ranking constraint loss.

[0041] Step S3 also includes: Based on the ensemble learning model loss function, the weight parameter update value is outputted by the training set according to the actual load value in the historical prior data and the load quantile prediction value.

[0042] The load quantile forecast value is obtained based on the forecast value obtained by the meteorological forecast value, holiday information, etc. in the historical prior data. The loss value is calculated according to the difference between the actual load value and the forecast value to obtain the weight parameter update value, so as to update the weight parameter of each sub-model in the ensemble learning model, thereby improving the output accuracy of the ensemble learning model. By combining the predictions of multiple sub-models, the ensemble learning model can reduce the bias and / or variance of a single model, thereby improving the accuracy of the overall prediction. In this process, the load quantile forecast value and the actual load value are obtained through the historical prior data in the validation set. Since different models may be sensitive to different aspects of the data, it can provide the robustness of the overall prediction. By dynamically adjusting the weights of the sub-models through the ensemble learning model loss function and the mixer, the ensemble learning model is better adapted to different data sets and prediction tasks.

[0043] In this embodiment, outputting a continuous probability density curve according to a set of load quantile prediction values ​​based on an adaptive bandwidth kernel density estimation algorithm also includes: Calculate the local density under the initial bandwidth for each load quantile prediction value in the load quantile prediction value set; The geometric mean density is calculated using the local densities corresponding to all load quantile prediction values; The adaptive bandwidth of each load quantile prediction value is calculated using the geometric mean density; The adaptive bandwidth is used to calculate the probability density estimate of the load under the time series to be predicted, and a continuous probability density curve is output.

[0044] In this embodiment, the ensemble learning model outputs m load quantile prediction values ​​at time i: At this time, m is 99, that is, the adaptive bandwidth kernel density estimation algorithm is used to convert the discrete 99 load quantile prediction values ​​into a continuous probability density curve.

[0045] The adaptive bandwidth method dynamically adjusts the smoothing parameters according to the data characteristics, provides an adaptive smoothing effect, and better reflects the real change pattern of the load. The load probability density at time i is obtained through adaptive bandwidth kernel density estimation:

[0046] 1) Initial local density estimation: For each quantile Calculate the local density under the initial bandwidth h Among them, q j represents the j-th quantile, j = 1, 2, 3, ..., m; K(u) is the Gaussian kernel function; h is the initial bandwidth.

[0047] 2) Calculate the geometric mean density of all load quantile prediction values

[0048] 3) Calculate the adaptive bandwidth of each load quantile forecast value Among them, α is 0.5.

[0049] 4) Adaptive Kernel Density Estimation: Use adaptive bandwidth Calculate the probability density estimate of the load value y at time i:

[0050] The kernel density estimation algorithm with adaptive bandwidth can dynamically adjust the bandwidth according to the local characteristics of the data, thereby providing a more accurate probability density estimation.

[0051] As the second embodiment of the present application, a specific embodiment is provided to prove it: the power consumption data of a 10kV bus load is collected, and the load data collection time length is one year, and the collection interval is 15min. The meteorological data includes ultraviolet radiation exposure index, temperature, humidity and air pressure, etc. The collection interval is 1h. In order to ensure that the dimensions of meteorological data and load data are consistent, linear interpolation is used to interpolate and fill the load data. The models used are all built based on Keras and Tensorflow deep learning platforms in Python.

[0052] The comparison model uses the parametric model mixed density network (MDN) model and the QRNN model in the non-parametric model, and the improved algorithm QRGRU model. In order to verify the improvement effect of the sorting algorithm and the integrated model on the multi-value quantile regression prediction, the comparison model type and parameter settings are set as shown in Table 1. The last month's data is used as test data, including holidays, weekdays and weekends, and the rest of the data is used as training verification data. The ratio of training data to verification data is 7:3. The prediction form of each model predicts the load quantile value of 0.01-0.99 quantiles under a single time step in advance, and further uses the adaptive bandwidth kernel density estimation to obtain the probability density and cumulative probability distribution of the load change, and then calculates the predicted value at a specific probability level, so as to estimate the load value interval.

[0053] The optimization algorithm uses the Adam algorithm. In order to make the model reach the optimal state during the training process, the total number of model iterations is set to 100, and the automatic termination algorithm is used. The initial learning rate of the model is set to 0.001, and the exponential function decay scheduling mode is used to deal with the situation where the loss function does not decrease during the training process. When the loss function no longer decreases, the learning rate is automatically reduced to continue to optimize the model at a smaller learning rate to ensure that the model can fully converge and obtain the best performance.

[0054] Table 1 Comparison of model types and parameter settings

[0055] After model training and parameter adjustment based on the training data set, the multi-value quantile regression based on the maximum likelihood estimation sorting algorithm significantly reduced the cross-degree of multi-value quantiles, which was 53.786% less than the traditional quantile regression model. Therefore, the method proposed in this application can reduce the cross-degree during load multi-quantile numerical prediction.

[0056] The integrated learning load probability prediction model constructed in this application has better solved the multi-objective problem corresponding to multi-value quantile regression, reduced the interval width of the load probability prediction, and also obtained a more accurate interval coverage rate; by comparing the load probability prediction examples of different day types, the deviation of the probability prediction interval coverage rate was reduced, and the improvement effect on the load probability prediction of weekdays and weekends with large fluctuations is more significant.

[0057] In summary, the load probability forecasting method proposed in this application has high reliability and sharpness, and can provide an accurate data basis for power system operation decisions.

[0058] The specific implementation described above is a preferred implementation of the short-term probability load forecasting method based on multi-value quantile regression and integrated learning of the present application, and is not intended to limit the specific implementation scope of the present application. The scope of the present application includes but is not limited to the specific implementation embodiment. All equivalent changes made in accordance with the shape and structure of the present application are within the protection scope of the present application.

Claims

1. A short-term probability load forecasting method based on multi-value quantile regression and ensemble learning, characterized by: The steps include: S1: Obtain historical load data, use the historical load data to divide it into historical benchmark data and historical prior data according to the time series, use the historical benchmark data to build several ensemble learning sub-models, and use the historical prior data and several ensemble learning sub-models to obtain the load quantile prediction value training set corresponding to the historical prior data; S2: Perform initial dynamic weighting on several ensemble learning sub-models to construct an ensemble learning initial model; S3: Based on the loss function of the ensemble learning model, the weight parameter update value is output according to the historical prior data and the load quantile forecast value training set, and the weight parameter update value is used to perform secondary dynamic weighting, update the ensemble learning initial model, and obtain the ensemble learning model; S4: Obtain the time series data to be predicted, and output a set of load quantile prediction values ​​corresponding to the time series to be predicted using the time series data to be predicted and the integrated learning model; S5: Based on the adaptive bandwidth kernel density estimation algorithm, a continuous probability density curve is output according to the load quantile prediction value set to obtain the load short-term probability prediction result.

2. The method for short-term probability load forecasting based on multi-value quantile regression and ensemble learning according to claim 1 is characterized in that: The method of constructing a plurality of integrated learning sub-models using historical benchmark data further includes: The historical benchmark data is randomly divided into several training data subsets, and a corresponding ensemble learning sub-model is constructed with each training data subset to obtain several ensemble learning sub-models.

3. The method for short-term probability load forecasting based on multi-value quantile regression and ensemble learning according to claim 2 is characterized in that: The constructing of a corresponding integrated learning sub-model with each training data subset also includes: Construct an integrated learning basic sub-model including feature selection network, LSTM encoding layer, LSTM decoding layer and gated residual network; Use feature selection networks to identify influential feature factors in a subset of training data; The LSTM encoding layer is used to obtain the short-term load change characteristics based on the influencing characteristic factors; An integrated learning sub-model is constructed using the short-term load variation characteristics as the input of the LSTM decoding layer and the gated residual network.

4. The method for short-term probability load forecasting based on multi-value quantile regression and ensemble learning according to claim 1 is characterized in that: The output of the weight parameter update value based on the integrated learning model loss function according to the historical prior data and the load quantile forecast value training set also includes: The loss function of the integrated learning model is constructed using the multi-value quantile loss function and the maximum likelihood estimation ranking loss.

5. The method for short-term probability load forecasting based on multi-value quantile regression and ensemble learning according to claim 4 is characterized in that: The method of constructing an integrated learning model loss function using a multi-valued quantile loss function and a maximum likelihood estimation sorting loss also includes: constructing a multi-valued quantile regression model based on a pinball loss function combined with quantile regression and a neural network model, and obtaining a multi-valued quantile regression loss function using the multi-valued quantile regression model.

6. The method for short-term probability load forecasting based on multi-value quantile regression and ensemble learning according to claim 1 is characterized in that: The method of outputting a continuous probability density curve according to a set of load quantile prediction values ​​based on an adaptive bandwidth kernel density estimation algorithm also includes: Calculate the local density under the initial bandwidth for each load quantile prediction value in the load quantile prediction value set; The geometric mean density is calculated using the local densities corresponding to all load quantile prediction values; The adaptive bandwidth of each load quantile prediction value is calculated using the geometric mean density; The adaptive bandwidth is used to calculate the probability density estimate of the load under the time series to be predicted, and a continuous probability density curve is output.

7. The method for short-term probability load forecasting based on multi-value quantile regression and ensemble learning according to claim 1 is characterized in that: The use of historical load data divided into historical benchmark data and historical advance data according to time series includes: The historical load data is divided into several time segments according to the preset time segments, with odd-numbered time segments as historical benchmark time segments and even-numbered time segments as historical prior time segments, and historical benchmark data and historical prior data are obtained based on the historical benchmark time segments and the historical prior time segments.

8. The method for short-term probability load forecasting based on multi-value quantile regression and ensemble learning according to claim 1 is characterized in that: The S2 further includes: The mixer is used to perform initial dynamic weighting on several ensemble learning sub-models to construct an ensemble learning initial model.

9. The method for short-term probability load forecasting based on multi-value quantile regression and ensemble learning according to claim 5, characterized in that: The multi-value quantile regression loss function is: Among them, ML(y i x i ) is the condition input x i Time i The multi-valued quantile regression loss function, y i is a vector of several load quantile forecast values ​​at time i, m is the total number of quantiles, q j represents the j-th quantile, The quantile is q j The loss function when , g(xiqj) is the conditional input xi, quantile q j Multi-value quantile prediction model for .

10. The method for short-term probability load forecasting based on multi-value quantile regression and ensemble learning according to claim 1, characterized in that: The historical load data at least includes historical meteorological data, historical time series data and historical load data.

Citation Information

Cited By

  • Industrial park net load prediction method and system based on multi-layer perceptron

    CN119398279A

  • Probability load prediction method and system for urban and rural residents

    CN119891151A

  • Electric quantity prediction method based on small sample data

    CN120031264A

  • A method for power consumption prediction based on small sample data

    CN120031264B

  • Training method of energy consumption probability prediction model, energy consumption probability prediction method, computer equipment and readable storage medium

    CN120596908A